Statistical Versus Simulation Models in Safety

Steps Toward a Synthesis Using Median-Crossing Crashes

Gary A. Davis and Paul Morris
Rational planning involves selecting from possible courses of action with regard to their predicted consequences. In road safety engineering, this requires the ability to predict the frequency and severity of crashes that are expected to result from design or operational changes. Statistical modeling is the dominant approach used in the rst edition of the Highway Safety Manual to address this issue, but over time statistical models are expected to be replaced by models that explicitly describe mechanisms underlying crash occurrence. This development could be accelerated if the already substantial investment in statistical safety models could be used to calibrate or validate mechanism-based simulation models. This paper briey reviews two efforts in which statistical models of mediancrossing crash (MCC) frequency were developed. An objective of both statistical studies was to summarize how the expected frequency of MCCs varies with the average daily traffic and median width of a divided highway. After the nature of the simulation model is described, the summarized relationships are used as criteria for assessing its performance.

Rational planning involves selecting from possible courses of action with regard to their predicted consequences. In road safety engineering, this requires being able to predict the frequency and severity of crashes that are expected to result from design or operational changes. Hauer (1) has argued persuasively for developing a scientically justiable methodology for making these predictions, and the Highway Safety Manual (HSM) is a clear effort to address this issue, at least partially. The dominant approach used in the rst edition of the HSM is statistical modeling of observational data, where generalized linear models are used to describe baseline associations between crash frequency and observable road features, while the effect of interventions is captured through empirically determined crash modication factors. The effect of an intervention is predicted by using the base model to forecast crash frequency under the prevailing conditions and then multiplying the expected frequency by the crash modication factor. Ideally, the crash modication factor was estimated from a wellconducted beforeafter study that controlled for selection bias effects. In practice, though, this is not always the case, and provisional estimates of crash modication factors have been determined by expert review of methodologically weaker work. A substantial research
G. A. Davis, Department of Civil Engineering, University of Minnesota, 500 Pillsbury Drive SE, Minneapolis, MN 55455. P. Morris, SRF Consulting Group, Inc., One Carlson Parkway North, Suite 150, Minneapolis, MN 55447. Corresponding author: G. A. Davis, drtrips@umn.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2102, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp. 93100. DOI: 10.3141/2102-12

effort, taking several years and costing several million dollars, has supported the rst edition of the HSM. In a presentation (Theory, Explanation, and Prediction in Road Safety) to a 2006 TRB workshop, Bonneson and Lord pointed out an interesting analogy with the development of the Highway Capacity Manual, in which regression models for predicting traffic signal delay by using nave specications of independent variables were replaced by regression models in which the form of the independent variables was justied theoretically. Those models in turn were replaced by models in which the functional form relating traffic ow, capacity, and signal timing to delay was justied theoretically. A similar evolution may be expected for the HSM, in which rst edition regression models are gradually replaced by structural models that explicitly describe mechanisms underlying crash occurrence. Although causal inference is possible in the absence of plausible mechanisms, when an observed association can be explained by a plausible or known mechanism, the claim of a causal relationship is reinforced (2). And as experience with the Highway Capacity Manual shows, simulation models that capture underlying mechanisms are usually able to represent a richer and more detailed set of alternatives than are statistical models. Although a research commitment to structural modeling similar to that devoted to statistical modeling would be welcome, it should probably not be expected until structural modeling can achieve a similar focus on wellspecied research questions. In the meantime, this evolution can be accelerated if the substantial investment in statistical models can be used to support the development of structural models. This paper will illustrate how published statistical models might be used to support initial calibration and validation of a model simulating the frequency of median-crossing crashes (MCCs) on divided highways. The basic idea is that since statistical models provide defensible summaries of the associations between crash frequency and measured roadway features, a plausible simulation model should be able to reproduce those associations, at least within the uncertainty inherent in the statistical models themselves. Two efforts in which statistical models of MCC frequency were developed are reviewed briey. An objective of both statistical studies was to summarize how the expected frequency of MCCs varied with the average daily traffic (ADT) and the median width of a divided highway. After the nature of the simulation model is described, the summarized relationships are used as criteria for assessing the performance of the simulation model.

STATISTICAL MODELING OF MCCs The study by Donnell et al. (3) focused on cross-median crashes on Pennsylvania freeways and expressways. Computerized crash records and road inventory data for 19941998 were compiled,
93

94

Transportation Research Record 2102

with totals of about 2,500 mi for freeways and 1,400 mi for expressways. Since Pennsylvanias computerized crash records did not identify MCCs explicitly, hard copies of crash reports for candidate crashes were studied. A total of 267 cross-median crashes were identied, 52% of which were on Interstate highways and 48% on expressways. Generalized linear modeling was used to relate the ADT and median width of a road section to the expected number of MCCs. The model for Interstate highways took the following form: N MCC = exp ( a + cW ) ( L ) ( ADT ) where NMCC L ADT W a, b, c = = = = = expected number of MCCs per year per direction; section length (mi); directional average daily traffic (vehicles per day); median width (ft); and model parameters.
b

Both models identify traffic volume and median width as important predictors of MCC frequency. The Texas study included the number of lanes and a speed limit effect in its model and appears to have detected a time trend in MCC frequency. Note that in the Pennsylvania model MCC frequency varies as a quadratic function of ADT, while in the Texas model the relation is linear. The models can differ by as much as a factor of 2 with respect to the MCC frequencies predicted for approximately similar conditions.

(1)

SIMULATION MODEL FOR MCCs Arguably, the best-developed simulation model for roadway safety is the Roadside Safety Analysis Program (RSAP) (5, 6). However, RSAP is designed to model collisions with roadside objects and does not explicitly support modeling MCCs. The MCC simulation model presented here is similar to RSAP in spirit but more modest in scope. The fundamental equation of the model is E [C ] = E [ N ] P [U N ] P [ X U ] P [C X ] where E[C] E[N] P[U N] P[X U] expected number of MCCs per year, expected number of median encroachments per year, probability that an encroachment is uncontrolled, probability that an uncontrolled encroachment crosses the median, and P[C X] = probability that a crossing encroachment collides with an opposing vehicle. = = = = (3)

A subsequent analysis (E. Donnell, personal communication, 2008) gave the updated point estimates and approximate standard errors for the model parameters listed below.
Parameter a b c Point Estimate 21.628 2.044 0.26 Standard Error 1.833 0.185 0.006

As part of the trend toward reevaluating practices with regard to median barriers, the Texas Department of Transportation funded a team at the Texas Transportation Institute to investigate this issue (4). The research team identied 52 Texas counties where a preponderance of potential MCCs occurred and requested hard copies of crash reports for 791 potential MCCs for 19981999. Inspection of the reports identied 443 apparently genuine MCCs. Roadway inventory data were also compiled, but data limitations prevented separate directional analyses. Generalized linear modeling was used to relate MCC frequency to measured site features, resulting in models of the following form: 365 ( ADT ) ( L ) N MCC = 1, 000, 000 exp [ a + b ( year ) + cW + d ( nlane ) + e ( speed )] where NMCC ADT year nlane speed = = = = = expected number of MCCs per year, both directions; total (not directional) ADT (vehicles per day); 0 for 1998, 1 for 1999; number of lanes; and 1.0 for 65-mph speed limit, 0 otherwise. (2)

Encroachment The expected number of encroachments on a section of highway is given by the following relation: E [ N ] = ( MVMT ) (4)

L and W are as defined previously; a, b, c, d, and e are model parameters. The table below gives estimates for the parameters appearing in Equation 2.
Parameter a b c d e Point Estimate 3.779 1.163 0.011 0.293 0.5 Standard Error 0.48 0.14 0.003 0.09 0.16

where MVMT is million vehicle miles of travel on the section and is encroachments per MVMT. MVMT is readily determined once the length and the ADT of the section are known, but determining the encroachment rate is more difficult. A median encroachment occurs when a vehicle leaves the traveled way and enters the median separating the directional lanes of a divided highway. The encroaching vehicle may come to a stop in the median, may return to its original traveled way without stopping, or may traverse the median and enter the traveled way of the opposing traffic. Ideally, estimates of the rate of encroachment and the proportion of encroachments of different types would be obtained by direct observation of the road in question over a suitable period of time. Unfortunately, the cost and time needed to collect such data are prohibitive, and the developers of encroachment models have had to rely on a limited number of studies carried out in the past. Hutchinson and Kennedy (7 ) conducted a study of median encroachments on a 24.6-mi-long section of Interstate 74 in Illinois. In that study, research teams periodically drove both directions of the freeway looking for wheel tracks in the freeways median. When they were found, the team measured the directional angle of the encroaching vehicles path and the extent of its lateral and longitudinal travel if it appeared to stop in the median, or, if it crossed the

Davis and Morris

95

median, the longitudinal extent of the point where it left the median. No measurements of a vehicles path were made after it left the median. The observers also classied each track as to whether it was essentially straight-line track or there was evidence of steering control. The study ran from October 1960 to April 1964, during which slightly more than 300 median encroachments were recorded. The other major encroachment study was conducted by Cooper (8, 9) in the late 1970s on a sample of highway sections in ve Canadian provinces. Both median and right-hand encroachments were recorded, and the sample included a range of roadway functional classications. After a review of the reports by Hutchinson and Kennedy and Cooper, the decision was made to rely on the data collected by Hutchinson and Kennedy. This was primarily because the Hutchinson and Kennedy study focused on median encroachments on a high-speed divided highway. Hutchinson and Kennedys report tabulated the data from each of their observed encroachments, and from that table a count of the number of encroachments recorded for each month of the study was prepared, along with estimates of the ADT for each month. A statistical analysis of encroachment rate versus time revealed distinctly higher encroachment rates during the rst half of the study than in the second half and an increase in encroachment rate during the months of December, January, and February. Because the start of Hutchinson and Kennedys study coincided with the opening of the freeway section, the initial high observed encroachment rate was thought to be a transient effect, and the decision was made to rely on the estimates from the last portion of the study. This gave the following estimated encroachment rates:
Time Period MarchNovember DecemberFebruary Encroachments per MVMT 0.77 1.20 Approximate Standard Error 0.12 0.25

input is a description of the medians cross section, consisting of, for each major break in the medians cross slope, the slope, the lateral width, and the friction coefficient for the slopes surface. Given such a cross-section description and initial values for v, , and b, the median traversal model proceeds as follows: 1. The deceleration for the cross-section segment is computed from b and the slopes friction factor, and decelerations in the lateral (x) and longitudinal (y) directions are computed on the basis of this, the angle , and the segments cross slope. 2. The lateral distance needed to bring the vehicle to a stop is computed. If this is less than the lateral extent of the segment, the stopping position of the vehicle is computed, and the encroachment is coded as one where the vehicle stops in the median. 3. If the lateral component of the vehicles stopping distance exceeds the lateral extent of the segment, the width of the segment (x) is added to the total lateral distance traveled, the corresponding longitudinal distance y = x/tan() is computed and added to the total longitudinal distance traveled, and the reduction in the vehicles kinetic energy is computed and subtracted from the total. 4. The above computations are then repeated for the next crosssection segment. 5. If the vehicle reaches the far side of the median without stopping, , b, and its speed on exiting the median become inputs into the collision simulation routine.

Collision When an encroaching vehicle reaches the far side of the median, its trajectory is modeled by numerically solving a system of ordinary differential equations, one equation each for its lateral (x) and longitudinal (y) position, one each for its speeds in the x and y directions, and one each for its x and y decelerations. The total deceleration at the start of rst opposing lane is computed as in the median traversal routine and decomposed into x- and y-direction components by using , and these are treated as constant throughout the vehicles trajectory across the opposing lanes. The initial speed is also decomposed into x and y components, and then for each small time step, the position and speed of the vehicle are updated by using a simple Eulers method. In addition, when an encroaching vehicle reaches the far side of the median, two opposing vehicles are generated for each lane on the opposing side of the highway. An initial location and speed are assigned to each of the opposing vehicles, along with a reaction time for the opposing driver and an emergency braking deceleration. In each lane, the initial location of the rst opposing vehicle is generated as a random outcome from an exponential random variable with mean equal to the reciprocal of the user-dened traffic density, while the initial location of the second vehicle in a lane is the sum of the rst vehicles location and an additional exponential random outcome. These then dene initial conditions for a set of coupled equations similar to those describing the encroaching vehicles trajectory, with each opposing vehicles longitudinal deceleration being zero until that drivers reaction time is exceeded, after which the longitudinal deceleration is set to the emergency value. For each time step, the location and speed of the opposing vehicles are updated with the simple Eulers method, and the distance between each opposing vehicle and the encroaching vehicle is computed. If the distance is less than a user-specied collision threshold, a collision is recorded.

Documentation of this initial work is given elsewhere (10).

Median Traversal As indicated above, Hutchinson and Kennedys report listed data for each of their observed encroachments. One of the features that the observers looked for was whether the wheel tracks indicated steering by the driver, as evidenced by curving tracks. Fortunately, the majority of median encroachments in Hutchinson and Kennedys data set were those where no steering was evidenced (i.e., where the vehicle path was approximately linear). As long as the focus was on encroachments that could result in collisions with opposing traffic, little was lost by restricting attention to this subset. An inspection of the encroachment list in Hutchinson and Kennedys report indicated that 108 out of 308 encroachments were classied as showing no steering control, so the estimate of P[U N] for this study was P [U N ] = 108 308 = .351 (5)

Since the encroaching vehicles trajectory is modeled as a straight line, the only vehicle-specic inputs required are its initial speed (v), its trajectory angle with respect to the roadway (), and its deceleration. Deceleration was modeled by specifying a braking factor, denoted by b, which is the fraction of the frictional deceleration used. For example, if the friction coefficient for the surface being traversed is 0.4 and the braking factor b = 0.5, then the deceleration of the vehicle would be (0.5)(0.4)g = 6.44 ft/sec2. The other required

96

Transportation Research Record 2102

Otherwise, if the encroaching vehicle comes to a stop or crosses all opposing lanes without colliding, a noncollision is recorded.

MODEL APPLICATION Computing Expected Crash Frequencies

output of the routine is consistent with the lateral and longitudinal measurements from the sampled encroachment, the combination of , v, and b is retained. Otherwise, new values for v and b are sampled and the process repeated until a combination of values for , b, and v is obtained that reproduces the observed encroachment.

Model Implementation and Calibration On the basis of the basic Equation 3, computation of the expected number of MCCs on a highway section requires computing the probability that an encroaching vehicle crosses the median, P[X N], and the probability that a crossing vehicle collides with an opposing vehicle, P[C X]. The above-described trajectory model is deterministic in that once the appropriate initial conditions are assigned to the encroaching and opposing vehicles, a collision either occurs or does not occur. The random element comes from placing a probability distribution over the set of possible initial conditions. That is, each combination of initial conditions is assigned a probability, and the probability of a collision is then the probability assigned to all combinations of initial conditions that result in a collision. In principle, the desired probabilities can be computed by integrating this probability distribution over the appropriate subset, but the complexity of the model relating the initial conditions to the outcome means that closed-form solutions to this integration problem are not to be expected. Numerical integration via quadrature would be a possibility. However, for example for two opposing lanes, a total of 19 values (three for the encroaching vehicle and four for each opposing vehicle) need to be specied, and the relative high dimensionality of this problem makes the application of quadraturebased methods problematic. The remaining option is Monte Carlo integration, under which a computer is used to generate a large pseudorandom sample of initial conditions, the deterministic model is applied to each of them, and the relative frequency of simulated collisions is recorded. As the size of the sample becomes arbitrarily large, the law of large numbers implies that the relative frequency will converge to the true probability. This is the tactic adopted here. For each opposing vehicle, the initial speed is sampled from a normal distribution with mean and standard deviation specified by the user. The initial locations of the opposing vehicles are sampled from an exponential distribution governed by a user-specied traffic density. The reaction times and braking decelerations of the opposing drivers are sampled from normal distributions with means and standard deviations derived from the emergency braking study conducted by Fambro et al. (11). The main challenge was determining an appropriate distribution for the initial speed, direction, and braking factor for the encroaching vehicle. The decision was to generate them as random outcomes consistent with the data collected by Hutchinson and Kennedy. To this end, 97 encroachments were identified where no evidence of steering was observed, where the vehicle was not stopped by a collision with a fixed object, and where the encroaching angle and the lateral and longitudinal distances were recorded. The sampling algorithm then proceeds as follows. One of the 97 observed encroachments is sampled at random from the list, and the encroaching angle is set to that observed value. A random speed is then generated from a uniform distribution with bounds 10 and 70 mph, and a random braking factor is generated from a uniform distribution with bounds 0.1 and 1.0. These then serve as inputs to the mediancrossing routine, with a cross-section description corresponding to the typical cross section listed by Hutchinson and Kennedy. If the The simulation model was implemented as an Excel spreadsheet, with the Monte Carlo simulation being carried out by a specially written macro. Details of the implementation can be found elsewhere (12). The simulation model contains a number of user inputs, all of which inuence the models predictions. Many of the inputs (e.g., cross-section specication, ADT, mean and standard deviation of opposing traffic speeds) describe road and traffic conditions that can be adjusted to reect the situation on the road section of interest. The inputs describing the behavior of the opposing drivers, means and standard deviations for the reaction time and braking deceleration, are taken from the statistics reported by Fambro et al. (11). One input that needs more explanation, however, is the crash closeness threshold. The simulation model uses this input to determine when an encounter between an encroaching vehicle and an opposing vehicle results in a crash, and its necessity is a consequence of modeling vehicles as particles rather than as rigid bodies. The models results are sensitive to this parameter, and it was desirable to ensure that a reasonable value was used. To this end, the value of the crash threshold was adjusted to approximate the output of the Pennsylvania model described earlier. A freeway section 1 mi long, with a median width of 60 ft and two lanes for each direction, was selected for calibration purposes. The initial speeds of opposing vehicles were taken to be normally distributed with a mean of 65 mph and a standard deviation of 5 mph. The median width was chosen because the majority of sites in the Pennsylvania data set had median widths greater than 50 ft, and it was believed that the Pennsylvania model more accurately reects crash tendencies on sections with wider medians. The updated Pennsylvania model was then used to compute expected MCC frequencies for directional ADTs ranging from 5,000 to 40,000, and they are displayed in Table 1. The simulation model was then used to compute similar predicted crash frequencies, with the crash threshold being varied between 3.0 and 6.0 ft, to determine which crash threshold gives predictions most closely matching those of the Pennsylvania model. The results for crash thresholds of 4.0, 4.5, and 5.0 ft are also shown in Table 1, along with the percentage differences between the Pennsylvania model and the simulation model. For a crash threshold of 4.5 ft, the percentage difference for these Monte Carlo runs did not exceed 22%, while for shorter or longer crash thresholds the percentage difference could exceed 30% (see bold entries in Table 1).

Initial Validation One application of the statistical models described above is to compute predicted frequencies of MCCs as a function of median width and ADT. If the statistical models are accepted as reasonable summaries of the associations between MCC frequency, ADT, and median width, the simulation model should generate predictions that are roughly consistent with the statistical models. A set of

Davis and Morris

97

TABLE 1

Calibration of Simulation Model to Pennsylvania Model, Median Width Predicted Median Crossing Crashes per Direction per Year

60 ft

Directional ADT 5,000 10,000 15,000 20,000 25,000 30,000 35,000 40,000

Pennsylvania .0031 .0128 .0292 .0526 .083 .1205 .1651 .2169

Simulation, 4.0 ft .0032 (4) .0118 (8) .0250 (14) .0464 (12) .0672 (19) .0831 (31) .1184 (28) .1551 (28)

Simulation, 4.5 ft .0038 (22) .0149 (17) .0306 (5) .0497 (5) .0772 (7) .0996 (17) .1431 (13) .1703 (22)

Simulation, 5.0 ft .0040 (30) .0157 (23) .0363 (24) .0614 (17) .0883 (6) .1233 (2) .1670 (1) .1962 (10)

representative cross sections consistent with Minnesota Department of Transportation design guidelines, with median widths ranging from 30 to 70 ft, was prepared, and the simulation model was used to compute predicted MCC frequencies. The results are summarized in Table 2. The predictions given in Table 2 were then compared with similar predictions made by the Pennsylvania model and the Texas model. The comparison measure was the following t-statistic: t= ln ( mstat ) ln ( msim ) sstat (6)

rithms of the predicted frequencies were used primarily because, as Table 2 shows, the predicted MCC frequencies can vary by two orders of magnitude over the ranges of median width and ADT. Since both the Pennsylvania and Texas models are examples of log-linear models, with the form mstat = exp i xi i

(7)

where mstat = predicted MCC frequency from the statistical model, msim = predicted MCC frequency from the simulation model, and sstat = standard deviation for ln(mstat) due to uncertainty in model parameters. A justication for this measure is as follows. For large enough samples, central limit theorem results imply that the uncertainty in the coefficients of a generalized linear model can be approximated by using a multivariate normal distribution. The natural logarithm of the generalized linear model is simply a linear combination of these coefficients, so its uncertainty would also be approximately normal. Given a null hypothesis that the expected frequency of the statistical model equals that of the simulation model and a large enough sample used in estimating the statistical models parameters, t as given in Equation 6 would be approximately distributed as a standard normal random variable. A t-statistic with absolute value greater than 1.6 could be regarded as indicating a signicant difference between the predictions generated by two models. The loga-

where xi is the value of independent variable i and i is the coefficient for independent variable i, an approximate standard deviation sstat is given by sstat =

(x s )
i i i

(8)

where si is the standard error of estimate for coefficient i. Table 3 shows the comparison between the simulation model and the Pennsylvania model. Within the uncertainty produced by the Pennsylvania models parameter estimates, the two models give similar results. Table 4 shows a similar comparison between the predictions generated by the simulation model and those produced by the Texas model. In this case, for a 30-ft median and directional ADTs of 30,000 or more, the simulation model predicts more MCCs than does the Texas model, while for an ADT of 5,000 and median widths of 60 or 70 ft, the simulation model predicts fewer MCCs than does the Texas model (see bold entries in Table 4). For all other median widthADT combinations, the two models are roughly equivalent.

TABLE 2

Simulation Models Predicted Median Crashes per Direction per Mile per Year Directional ADT (vehicles per day)

Median Width (ft) 30 40 50 60 70

5,000 .018 .012 .009 .003 .003

10,000 .070 .045 .036 .014 .012

15,000 .147 .098 .081 .033 .025

20,000 .249 .155 .121 .051 .042

25,000 .364 .245 .184 .074 .053

30,000 .512 .319 .248 .115 .092

35,000 .652 .409 .340 .132 .127

40,000 .828 .505 .427 .175 .142

98

Transportation Research Record 2102

TABLE 3

t-Statistic Differences Between Simulation Model and Pennsylvania Model Directional ADT (vehicles per day)

Median Width (ft) 30 40 50 60 70

5,000 0.40 0.33 0.34 0.04 0.04

10,000 0.37 0.29 0.31 0.04 0.09

15,000 0.33 0.27 0.30 0.05 0.04

20,000 0.30 0.21 0.22 0.01 0.01

25,000 0.27 0.21 0.20 0.04 0.07

30,000 0.25 0.17 0.17 0.02 0.00

35,000 0.22 0.14 0.17 0.08 0.00

40,000 0.21 0.12 0.15 0.08 0.07

TABLE 4

t-Statistic Differences Between Simulation Model and Texas Model Directional ADT (vehicles per day)

Median Width (ft) 30 40 50 60 70

5,000 0.66 1.18 1.37 2.70 2.90

10,000 0.38 0.15 0.32 1.58 1.59

15,000 0.92 0.45 0.32 0.89 1.14

20,000 1.30 0.71 0.50 0.66 0.78

25,000 1.55 1.08 0.80 0.43 0.76

30,000 1.80 1.21 0.98 0.04 0.21

35,000 1.94 1.36 1.24 0.07 0.04

40,000 2.11 1.49 1.38 0.16 0.01

Graphical displays of some of the results from Tables 3 and 4 are shown in Figures 1 and 2. Figure 1 compares the predictions from the simulation model and the Pennsylvania model, for a 40-ft median. Figure 2 shows a similar comparison between the simulation model and the Texas model, again for a 40-ft median. The primary structural difference between the Pennsylvania and Texas models is that in the Texas model, expected MCC fre-

quency is proportional to ADT, while in the Pennsylvania model, it is roughly proportional to the square of ADT. In the simulation model, expected encroachments are proportional to ADT in the original direction, while probability of collision is roughly proportional to ADT in the opposing direction. The overall result is a relation between ADT and MCC frequency that is more like that of the Pennsylvania model. This is clearly shown in Figure 2,

Log predicted MCCs

10

0 0

1104

2104 aadt j ADT

3104

4104

5104 50,000

Simulation model Pennsylvania model 1.6 standard deviations FIGURE 1 Log of predicted MCC frequencies produced by simulation model and Pennsylvania model for 40-ft median. Error bars indicate 1.6 standard deviation range for Pennsylvania model.

Davis and Morris

99

1
Log predicted MCCs

0 0

1104

2104 aadt j ADT

3104

4104

5104 50,000

Simulation model Texas model 1.6 standard deviations FIGURE 2 Log of predicted MCC frequencies produced by simulation model and Texas model for 40-ft median. Error bars indicate 1.6 standard deviation range for Texas model.

where the slope of the ADT versus expected MCC frequency relation for the simulation model is markedly steeper than that for the Texas model. Another difference between the Pennsylvania and Texas models is that the uncertainty range for the Pennsylvania model is substantially greater than that for the Texas model. This is in large part due to the standard error associated with the Pennsylvania models constant term. Even if the median width and ADT coefficients were known exactly, the uncertainty for this term would be (1.6)(1.833) = 2.9. In comparison, at ADT = 40,000 vehicles per day and median width equal to 70 ft, the sstat for the Pennsylvania model comes out to about 2.72, so that (1.6)(2.72) = 4.35. That is, more than half of the uncertainty range is due to the constant parameter.

SUMMARY AND CONCLUSION As indicated above, a substantial effort is under way to provide road safety engineers with predictive tools that should allow them to assess the effect of safety countermeasures rationally. A supporting research effort has been expended to produce regression-type statistical models relating crash experience to conveniently measured roadway features. However, over time, models explicitly describing mechanisms underlying crash occurrence should be expected to replace rst-edition regression models. This evolution may be accelerated if it can be leveraged with the substantial investment already made in statistical models. To this end, a description was given of how development of a simulation model for MCCs could be supported with published statistical models. The idea behind the approach is that if a statistical model provides a reasonable summary of the statistical associations in a data set, it should be possible to use the statistical model as a surrogate for the crash data used to develop that model. In this paper, (a) a statistical model was used to cali-

brate a simulation model by computing an estimate of a simulation model parameter, crash closeness, that approximately replicated the output of a simulation model, and (b) the predictions of the simulation model were compared with those of two statistical models to check the plausibility of the simulation models predictions. Not surprisingly, the simulation model best-matched a statistical model that allowed expected MCC frequency to increase approximately as a quadratic function of traffic volume, but over a substantial range of ADT and median width combinations, the differences between the simulation model and the statistical models were within the uncertainty ranges generated by the uncertainty in the statistical model parameters. At this point it may now be apparent that statistical models to some extent deemphasize limitations in existing data, while simulation models tend to highlight them, with sometimes disconcerting clarity. The simulation modeler will often then be denied the sense of completion and closure given by a well-calibrated and validated statistical model. Each of the main components of the MCC simulatorthe encroachment model, the median traversal model, and the collision modelwas constrained by limitations in available data, and each could benefit from more detailed model specification coupled with better-focused observations. This is regarded as a desirable feature of simulation models, not a bug. As in other areas of science, progress can be expected through an interactive evolution of targeted observations highlighting model weaknesses, leading to proposed improvements, followed by more focused observation.

ACKNOWLEDGMENT This research was supported in part by the Minnesota Department of Transportation.

100

Transportation Research Record 2102

REFERENCES
1. Hauer, E. The Engineering of Safety and the Safety of Engineering, In Challenging the Old Order: Towards New Directions in Traffic Safety Theory (J. Rothe, ed.), Transaction Publishers, London, 1990, pp. 2971. 2. Thagard, P. Explaining Disease: Correlations, Causes, and Mechanisms. Minds and Machines, Vol. 8, 1998, pp. 6178. 3. Donnell, E. T., D. W. Harwood, K. M. Bauer, J. M. Mason, Jr., and M. T. Pietrucha. Cross-Median Collisions on Pennsylvania Interstates and Expressways. In Transportation Research Record: Journal of the Transportation Research Board, No. 1784, Transportation Research Board of the National Academies, Washington, D.C., 2002, pp. 9199. 4. Bligh, R., S.-P. Miaou, D. Lord, and S. Cooner. Median Barrier Guidelines for Texas. Report 0-4254-1. Texas Department of Transportation, 2006. 5. Mak, K. K., and D. L. Sicking. NCHRP Report 492: Roadside Safety Analysis Program (RSAP)Engineers Manual. Transportation Research Board of the National Academies, Washington, D.C., 2003. 6. Mak, K. K., and D. L. Sicking. Roadside Safety Analysis Program (RSAP) Users Manual. TRB, National Research Council, Washington, D.C., 2002. http://onlinepubs.trb.org/Onlinepubs/nchrp/rsap_users_ manual.pdf.

7. Hutchinson, G., and T. Kennedy. Medians of Divided Highways: Frequency and Nature of Vehicle Encroachments. Bulletin 487. University of Illinois Engineering Experiment Station, Urbana, 1966. 8. Cooper, P. Analysis of Roadside Encroachments: Single Vehicle Runoff-Road Accident Data Analysis for Five Provinces. British Columbia Research Institute, British Columbia, Canada, 1980. 9. Cooper, P. Analysis of Roadside Encroachment Data from Five Provinces and Its Application to an Off-Road Vehicle Trajectory Model. British Columbia Research Institute, British Columbia, Canada, 1981. 10. Davis, G. A., and P. M. Morris. Reanalysis of Hutchinson and Kennedys Encroachment Frequency Data. In Transportation Research Record: Journal of the Transportation Research Board, No. 1953, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp. 6570. 11. Fambro, D. B., K. Fitzpatrick, and R. J. Koppa. NCHRP Report 400: Determination of Stopping Sight Distances. TRB, National Research Council, Washington, D.C., 1997. 12. Davis, G. Cross-Median Crashes: Identication and Countermeasures. Minnesota Department of Transportation, St. Paul, 2008.
All statements made here are strictly the responsibility of the authors and do not necessarily represent the views of the Minnesota Department of Transportation. The Safety Section sponsored publication of this paper.